Adaptive noise variance identification for data fusion using subspace-based technique

The Kalman filter is commonly employed to fuse the measured information from displacement and acceleration responses. This fusion technique can mitigate the noise effect and produce more accurate response estimates. The fusion performance however significantly depends on the accuracy of noise estimation. The noise variances are generally estimated empirically a priori and assumed to be fixed throughout the calculation. Hence some estimation error might occur when the noise characteristics are time-varying. In this study, an adaptive subspace-based technique is developed to identify the noise variances. The approach is based on the principal component analysis which decomposes noise-contaminated signals into the signal subspace and the noise subspace. The variances of the noise and signals can then be estimated independently. To track the time variations of the noise and signal variances in an on-line fashion, a projection approximation subspace tracking technique is employed. The proposed technique can be incorporated into an adaptive Kalman filter and provide a more accurate estimation for data fusion.